In biology, flaws are generally bad. However, in materials science, defects can be deliberately tuned to give materials modern, useful properties. Today, atomic-scale defects are carefully introduced into the manufacturing process of products like steel, semiconductors, and solar cells to improve strength, control electrical conductivity, optimize performance, and more.
However, even as defects have become a powerful tool, accurately measuring different types of defects and their concentrations in finished products remains a challenge, especially without cutting or damaging the final material. Without knowing what defects lie in their materials, engineers risk producing products that perform poorly or have unintended properties.
Now, MIT researchers have built an artificial intelligence model capable of classifying and quantifying certain defects based on data from a non-invasive neutron scattering technique. The model, which was trained on 2,000 different semiconductor materials, can detect up to six types of point defects in the material simultaneously, which would be impossible using conventional techniques alone.
“Existing techniques do not allow for accurate characterization of defects in a universal and quantitative manner without damaging the material,” says lead author Mouyang Cheng, a PhD student at the Faculty of Materials Science and Engineering. “With conventional techniques without machine learning, detecting six different defects is unthinkable. It’s something that can’t be done any other way.”
Scientists say the model is a step toward more correct apply of defects in products such as semiconductors, microelectronics, solar cells and battery materials.
“Right now, defect detection is like the saying about seeing an elephant: each technique only allows you to see part of it,” says senior author and associate professor of nuclear science and engineering Mingda Li. “Some people see the nose, others the trunk or ears. But it is extremely difficult to see the whole elephant. We need better ways to get a full picture of the defects because we need to understand them to make the materials more useful.”
Cheng and Li were joined on the paper by postdoctoral fellow Chu-Liang Fu, undergraduate researcher Bowen Yu, graduate student Eunbi Rha, graduate student Abhijatmedhi Chotrattanapituk ’21, and Oak Ridge National Laboratory staff members Douglas L Abernathy PhD ’93 and Yongqiang Cheng. The paper appears in the magazine today.
Defect detection
Manufacturers have mastered the ability to tune out defects in their materials, but measuring the exact amounts of defects in finished products still relies largely on guesswork.
“Engineers have many ways to introduce defects, such as through doping, but they still struggle with basic questions such as what type of defect they created and at what concentration,” Fu says. “Sometimes they also have undesirable defects, such as oxidation. They don’t always know whether they have introduced undesirable defects or impurities during the synthesis. This is a long-term challenge.”
As a result, there are often many defects in each material. Unfortunately, every method of understanding defects has its limitations. Techniques such as X-ray diffraction and positron annihilation characterize only certain types of defects. Raman spectroscopy can recognize the type of defect but cannot directly infer its concentration. Another technique, known as transmission electron microscopy, requires people to cut gaunt slices of samples for scanning.
In several previous papers, Li and colleagues applied machine learning to experimental spectroscopy data to characterize crystalline materials. In their modern work, they wanted to apply this technique to defects.
For the experiment, scientists created a computational database of 2,000 semiconductor materials. They created pairs of samples from each material, one doped for defects and the other left undefected, and then used a neutron scattering technique that measures the different vibrational frequencies of atoms in solid materials. Based on the results, they trained a machine learning model.
“This created a basic model of the 56 elements of the periodic table,” says Cheng. “The model uses a multi-head attention engine, similar to ChatGPT. It similarly extracts the difference in data between materials with and without defects, and then predicts what dopants were used and in what concentrations.”
The researchers refined their model, validated it against experimental data, and showed that it could measure defect concentrations in an alloy commonly used in electronics and in a separate superconducting material.
The researchers repeatedly doped the materials to introduce multiple point defects and test the limits of the model, ultimately finding that it could predict a maximum of six defects in the materials at once, with a defect concentration of just 0.2%.
“We were really surprised that it worked so well,” Cheng says. “Decoding mixed signals from two different types of defects – let alone six – is very difficult.”
A model approach
Typically, manufacturers of components such as semiconductors conduct invasive tests on a compact percentage of products coming off the production line. This is a ponderous process that limits their ability to detect any defect.
“Right now, people largely estimate the number of defects in their materials,” Yu says. “Checking estimates using each individual technique is a tedious experience that only offers local information in a single grain anyway. This creates confusion about what flaws people think they have in their material.”
The researchers found the results invigorating, but they note that their technique for measuring vibrational frequencies using neutrons would be hard for companies to quickly implement in their own quality control processes.
“This method is very effective, but its availability is limited,” says Rha. “Vibrational spectra are a simple idea, but in some setups they are very complicated. There are simpler experimental setups based on other approaches, such as Raman spectroscopy, that can be implemented more quickly.”
Li says companies have already expressed interest in the approach and asked when it would work with Raman spectroscopy, a widely used technique for measuring airy scattering. Li says the researchers’ next step will be to train a similar model based on Raman spectroscopy data. They also plan to extend their approach to detect objects larger than point defects, such as grains and dislocations.
For now, however, the researchers believe their study shows the inherent advantage of artificial intelligence techniques in interpreting defect data.
“To the human eye, these defect signals would look essentially the same,” Li says. “But AI’s pattern recognition is good enough to recognize different signals and get to the bottom of the truth. Defects are a double-edged sword. There are many good defects, but if there are too many of them, performance can suffer. This opens up a new paradigm in defect science.”
The work was supported in part by the Department of Energy and the National Science Foundation.